Available online at www.sciencedirect.

com

Journal of Banking & Finance 32 (2008) 347359 www.elsevier.com/locate/jbf

How do policy and information shocks impact co-movements of Chinas T-bond and stock markets?
Xiao-Ming Li *, Li-Ping Zou
Department of Commerce, Massey University (Albany), Private Bag 102 904, North Shore MSC, Auckland, New Zealand Received 9 November 2006; accepted 13 April 2007 Available online 16 July 2007

Abstract We investigate the impacts of policy and information shocks on the correlation of Chinas T-bond and stock returns, using originally the asymmetric dynamic conditional correlation (DCC) model that allows for the coexistence of opposite-signed asymmetries. The comovements of Chinas capital markets react to large macroeconomic policy shocks as evidenced by structural breaks in the correlation following the drastic 2004 macroeconomic austerity. We show that the T-bond market and the bondstock correlations bear more of the brunt of the macroeconomic contractions. We also nd that the bondstock correlations respond more strongly to joint negative than joint positive shocks, implying that investors tend to move both the T-bond and stock prices in the same direction when the two asset classes have been hit concurrently by bad news, but tend to shift funds from one asset class to the other when hit concurrently by good news. However, the stockstock correlation is found to increase for joint positive shocks, indicating that investors tend to herd more for joint bullish than joint bearish stock markets in Shanghai and Shenzhen. 2007 Elsevier B.V. All rights reserved.
JEL classication: G18; G10; C22; O53 Keywords: Macroeconomic austerity; Dynamic correlation; T-Bond market; Stock market; China

1. Introduction During 2003 and early 2004, China experienced an excessive investment boom. To cool this economic overheating, in AprilMay 2004 the government put into practice a series of tight policy measures. Included in these policy measures were the following. The central bank raised the reserve requirements and tightened credit lines. The China Banking Regulatory Commission required commercial banks to nix investment projects deemed to be illplanned, low quality, and unconformable to the governments industrial policies. The State Development and Reform Commission ordered local authorities to control the debut of price-hiking projects within their jurisdictions.
Corresponding author. Tel.: +64 9 414 0800x9471; fax: +64 9 441 8177. E-mail address: x.n.li@massey.ac.nz (X.-M. Li). 0378-4266/$ - see front matter 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.jbankn.2007.04.029
*

According to the news media, following the austerity programs, the Chinese stock and bond markets simultaneously underwent drastic drops, which subsequently had contagious eects on nancial markets in Hong Kong, the US, Japan, London, Australia and so on (for example, Japans stock price indexes reportedly fell by 400500 points). A metaphor went: As the Chinese economy is having an injection for allaying fever, the worlds nancial markets suer a shivering t (http://news.xinhuanet.com/fortune/ 2004-05/14/content_1468420.htm). These observations and anecdotes seem to suggest that, in China, drastic policy changes have begun to impact domestic nancial markets (as well as international nancial markets), which then motivated the present paper to attempt a serious investigation on some related issues into which anecdotes do not and cannot provide insights. However, we are not interested in how individual market returns, but rather how correlations between them, respond

348

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

to policy shocks. We have chosen to focus on the correlation between T-bond and stock returns for three main reasons as follows. First, to reduce portfolio risk via diversication, a key input required by risk managers to hold ecient portfolios is the correlation between assets included in the portfolio.1 Portfolios that contain stocks and government bonds have become popular among investors, as the two asset classes are believed to have dierent risk-return characteristics and their correlations to be low or even negative. Because the correlation between T-bonds and stocks plays a vital role in portfolio risk management and dynamic asset allocation for investors, it has been extensively studied in the literature. For example, an earlier study by Barsky (1989) looks at price co-movements between stocks and bonds, and nds that when investors are scared, they look for safety. They adjust their portfolios to include more safe assets and fewer risky assets. This kind of movement is usually referred to as a ight to quality. A recent study by Ilmanen (2003) on the US stockbond correlation reports that the correlation between stock market and government bond returns was positive through most of the 1900s, but negative in the early 1930s, the late 1950s, and recently. A negative correlation implies that investors have beneted from the bond market upswing, osetting some of their losses in stock markets. However, this combination may have severe implications for pension funding ratios, as both equities and discount rates decline, sending assets and liabilities in opposite directions. Second, correlations between the stock and bond markets are important to policymakers. Since China entered the WTO in 2000, the Chinese government has endeavored to reform its nancial system including capital markets, in order to transform the conduct of macroeconomic policies from being administrative to being market oriented in nature (as required by becoming a WTO member). In the latter context, the central bank cannot set specic price targets for stocks and bonds, and so has to utilize the information contained in the co-movements between the freely adjusted prices of these assets to gauge, for example, market participants expectations about growth and ination. In other words, the stockbond return correlation estimates may provide policymakers with useful complementary information to determine whether market participants are changing their views on ination or economic activity prospects. Quantifying contemporaneous relations between the stock and bond markets also helps policymakers to estimate and control the unintended consequences that policies directed primarily at one market could have for the other. To our best knowledge, the existing literature lacks such a study as ours for China, despite the important policy implications of the issues examined in the present paper.
An example that the correlation between three assets in a portfolio aects their optimal weights which maximize its Sharpe ratio is available from the authors upon request.
1

Third, correlations between asset returns have been viewed as an integral aspect of inter-nancial market integration, in the literature. Kim et al. (2006) and Berben and Jansen (2005) examine the dynamic or time-varying correlation between stock and T-bond returns of several European countries to infer the state and progress of their nancial integration, taking into account the inuence of the European Monetary Union as a possible cause of structural change. Kim et al. (2005) also conducted a similar study for stock market integration in Europe. In these studies, the authors use return correlations to gauge the degree of integration between nancial markets: a high/low correlation implies a high/low level of integration. High, not just low, stockbond correlations have also been established. For example, Kim et al. (2006) document that the interbondstock correlations for each of their sample Euro zone countries and the weighted average of these for Euro countries and also non-Euro zone countries once reached very high levels, although they have been falling since the mid 1990s. The authors take these results to imply a falling nancial integration since the mid 1990s. Apart from correlation analysis, the cointegration framework is also a useful tool in studying the degree of market integration, and a recent application of cointegration analysis to the longrun equilibrium relationship among Chinas money, stock and T-bond markets has appeared in Yin (2005). However, a long-run relationship detected by cointegration tests is credible only if the true relationship is constant or experiences few breaks over time. The fact that China has been reforming its nancial systems with frequent debuts of new reform programs allows one to argue that the number of breaks is too large to permit a meaningful cointegration analysis. Such an unstable, time-varying relationship among nancial markets ought to be modeled more appropriately within a time-varying correlation framework. By doing so, our investigation into the correlation between Chinas T-bond and stock returns will shed new light on inter-stockbond market integration that may vary across dierent points in time. However, we use the previously reported estimates of the European stockbond return correlations as a reference point with which to compare the estimates of the Chinese stockbond return correlations, in order to infer the relative degree of stockbond market integration in China. Our work contributes to the existing literature in at least two aspects. One is the link of stockbond correlations to information shocks or macroeconomic factors. Recent studies on these issues include Chordia et al. (2005) and Li (2002). In the former article that uses the US data, the authors nd that innovations to stock and bond market liquidity and volatility are signicantly correlated, and attribute this observation to the possibility that common factors such as monetary shocks and money ows drive liquidity and volatility in these markets. The latter paper shows that the major trends in stockbond correlation in G7 countries can be explained by their common exposure to macroeconomic factors, such as expected ination,

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

349

unexpected ination and the real interest rate. Our study focuses on China as the largest emerging economy, and links stockbond and stockstock correlations to macroeconomic austerity measures of an administrative nature and to the peculiar behavior of Chinese investors in response to information shocks. Exploring these unique characteristics of the Chinese nancial markets represents our attempt to ll the void in the literature. The other aspect is related to the technical front for empirically investigating time-varying return correlations. Scruggs and Glabadanidis (2003) introduces exibility into their specication for the time-varying covariance matrix of stock and bond returns, by assuming that conditional second moments follow an asymmetric dynamic covariance (ADC) process proposed by Kroner and Ng (1998). The use of the ADC model enables them to examine how return shocks and volatility are transmitted between the stock and bond markets. Cappiello et al. (2003) employ an asymmetric version of the dynamic conditional correlation (DCC) model which allows for a structural break to investigate asymmetric dynamics in the correlations of international equity and bond returns. They are able to nd strong evidence of asymmetries in conditional covariance of equity and bond returns, and signicant evidence of a structural break in conditional asset correlation upon the creation of the Euro. Connolly et al. (2005, 2007) estimate rolling correlations over time to examine whether the time-variation of stockbond and stockstock return comovements can be linked to stock implied volatility. It appears from the above-cited studies that those dierent approaches taken serve, and depend on, dierent objectives. Our research objectives (to be detailed below) determine that the asymmetric version of the DCC model with structural change as employed by Cappiello et al. (2003) seems to be a more appropriate econometric tool than those employed by Scruggs and Glabadanidis (2003) and Connolly et al. (2005, 2007). However, to adapt to the reality of the Chinese nancial markets, we modify the model employed by Cappiello et al. (2003) in the following innovative manners. First, we propose a new version of the model which is termed as the mixed asymmetric DCC (MADCC) model, and apply it for the rst time in the literature. Our MADCC model is capable of capturing concurrent responses, if any, of the conditional covariance of standardized return residuals to positive and negative information shocks. Second, we allow for multiple structural breaks in the conditional correlation. This enables us to test the strength and duration of policy impacts on the correlation. Third, we t a Skew-t-GARCH model (Jondeau and Rockinger, 2005) in ltering return volatility, so that the notorious problems of non-normality and skewness in the return distribution can be properly addressed. The objectives of this paper are twofold. One is to explore the general question of how the correlation responds to policy shocks. Specically, we ask: (1) In what direction did the 2004 macroeconomic austerities aect the stockstock and bondstock correlations? (2) How long

did the impacts last for? (3) Were there any dierences in the correlation patterns between the pre- and post-episode periods. (4) What are the possible implications of such differences for nancial market integrations, compared to the stockbond market integration in Europe as investigated in previous studies (e.g., Kim et al., 2006)? Answers to these questions should carry useful complementary information for policymaking. The second objective of this paper is to use the news impact surfaces of Kroner and Ng (1998) to study asymmetry in the impact of joint information shocks on correlation. Previous studies have only examined possible asymmetry in the reaction of univariate return volatility to information shocks for China, and evidence is mixed in terms of the sign of asymmetry (see, for example, Li, 2003a,b). We extend the literature and see whether correlation between two asset returns responds asymmetrically to joint bad news (represented by their last periods standardized return residuals being both negative) and joint good news (represented by their last periods standardized return residuals being both positive), and whether in the same way as developed economies nancial market correlations. Joint bad/good news means that two assets prices fall/rise together at a particular point in time. If, say, joint bad news (two prices falling together) increases the correlation between two assets, the probability that their prices will further fall together is high relative to the probability that they will further rise together following their previous rises together. This correlation asymmetry has important nancial implications. For example, the standard mean-variance investment theory advises portfolio diversication, but the value of this advice might be questioned if all the assets tend to fall as they have already fallen. Also, the presence of asymmetric correlations can potentially cause problems for hedging eectiveness, since hedging relies crucially on the correlation between assets hedged (and the nancial instruments used). For these reasons, asymmetry related to joint information shocks is of more concern than asymmetry related to mixed information shocks (i.e., two assets moving in the opposite directions). The paper is organized as follows. Section 2 describes the data used in this study, and provides their descriptive statistics. Econometric methods are presented in Section 3, followed by empirical results in Sections 4 and 5. Section 6 oers our conclusions. 2. Data Data used in this study include the daily Chinese Government Bond (T-bond) Price Index traded on the Shanghai Stock Exchange, the daily Shanghai Composite Price Index, and the daily Shenzhen Composite Price Index, for the period from January 2, 2003 to November 30, 2005. The T-bond price index data were obtained from the Security Industry Research Centre, Asia-Pacic (SIRCA), while data on the two equity indices were collected from Datastream. The Shanghai Stock Exchange

350

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359
Return rate (%, T-bond)
1.5 1 0.5 0 -0.5 -1 -1.5 -2
/01/0 3 2003 /02 2003 /24 /04 2003 /03 /05 2003 /22 /07/0 1 2003 /08 2003 /08 /09 2003 /17 /11/0 3 2003 /12 2004 /11 /02/0 2 2004 /03 2004 /11 /04/2 0 2004 /06 2004 /04 /07 2004 /14 /08/2 3 2004 /09 2004 /30 /11/1 6 2004 /12 2005 /24 /02/0 3 2005 /03 2005 /24 /05/1 0 2005 /06 2005 /17 /07/2 7 2005 /09 2005 /05 /10/2 0

(SSE) released for the rst time the Chinese T-bond price index on January 2, 2003. The T-bond price index, with a base value of 100, is the weighted average of all T-bonds traded on the SSE on the basis of their issuing volumes, and is the only index available for us to empirically investigate a T-bond market in China.2 The Shanghai and Shenzhen composite price index contain all traded stocks (A-shares and B-shares) on the two exchanges. They are considered to reect the overall price movements in the Chinese stock markets. Figs. 13 plot the (raw) return series of, respectively, the T-bond Price Index, the Shanghai Composite Price Index and the Shenzhen Composite Price Index, calculated by taking the rst dierence of the natural logarithm of the corresponding price indices. Table 1 presents the descriptive statistics for the three return series. The skewness statistic for the T-bond return series is negative and signicant at the 1% level, indicating that the return distribution is negatively skewed or there is a substantial probability of a big negative return. The skewness statistics for the Shanghai and Shenzhen stock index returns are both positive and signicant at the 1% level, implying that there is a substantial probability of a big positive return. The excess kurtosis statistics are all signicantly greater than 0 at the 1% level for the three return series, suggesting that the T-bond and stock return series are leptokurtic. The above statistics show that assuming a normal distribution for the three return series is inappropriate, which is conrmed by highly signicant JarqueBera statistics (at the 1% level) enabling one to reject the null hypothesis of normality. Regarding the LjungBox statistic, there is little formal guidance for the choice of the lag length, and so we decided to use a rule of thumb: the lag length is determined by the square root of the sample size as 26.3 With this lag length, the LjungBox Q statistics suggest that the two stock return series display no autocorrelation, as the statistics are not signicant at the 5% level. So when demeaning the two stock return series, no autoregressive terms need to be included. However, the T-bond return series suers a serious problem of autocorrelation, as the associated Q statistic is signicant at the 1% level. This prompts us to include autoregressive terms in demeaning the T-bond return series later for the DCC estimation. The ARCH test statistic is signicant at the 1% level for the T-bond, which suggests that the T-bond return series is heteroskedastic. The insignicant ARCH test statistics at the 5% level seem to indicate that the two stock return series are free of the ARCH problem. However, we should be cautious about these results, as their underlying normality assumption does not hold according to the normality test results just mentioned above. As will become clearer later,
Note that data on inter-bank bond markets are not externally available, and there does not yet exist a price index for all T-bonds traded on the Shenzhen Stock Exchange. 3 Slightly changing the lag length around 26 does not alter the test results.
2

2003

Fig. 1. Return rate of the T-bond price index.

Return rate (%, Shanghai stock)

10 8 6 4 2 0 -2 -4 -6
2003 /01/0 3 2003 /02 2003 /24 /04/0 3 2003 /05 2003 /22 /07/0 1 2003 /08 2003 /08 /09 2003 /17 /11/0 3 2003 /12 2004 /11 /02 2004 /02 /03 2004 /11 /04/2 0 2004 /06 2004 /04 /07/1 4 2004 /08/2 3 2004 /09 2004 /30 /11 2004 /16 /12 2005 /24 /02/0 3 2005 /03 2005 /24 /05 2005 /10 /06/1 7 2005 /07/2 7 2005 /09/0 5 2005 /10/2 0

Fig. 2. Return rate of the Shanghai stock price index.

Return rate (%, Shenzhen stock)

10 8 6 4 2 0 -2 -4
/01/0 3 2003 /02 2003 /24 /04/0 3 2003 /05 2003 /22 /07/0 1 2003 /08 2003 /08 /09 2003 /17 /11/0 3 2003 /12 2004 /11 /02 2004 /02 /03 2004 /11 /04/2 0 2004 /06 2004 /04 /07/1 4 2004 /08/2 3 2004 /09 2004 /30 /11 2004 /16 /12 2005 /24 /02/0 3 2005 /03 2005 /24 /05 2005 /10 /06/1 7 2005 /07/2 7 2005 /09/0 5 2005 /10/2 0

-6

2003

Fig. 3. Return rate of the Shenzhen stock price index.

under some non-Gaussian distribution assumptions, heteroskedasticity in the two stock returns can be detected. In summary, to capture possible skewness and excess kurtosis shown by the descriptive statistics, we will follow Jondeau and Rockinger (2005) by using the Skew-t-GARCH model to estimate the volatilities of the three return series.

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359 Table 1 Descriptive statistics Price index T-Bond Shanghai stock Shenzhen stock Skewness 1.8203** 0.7965** 0.5110** Kurtosis 14.928** 2.7095** 1.9275** Q(26) 174.66** 27.364 33.587 JB 6784.9** 283.44** 136.21** ARCH 156.01** 0.1025 0.0717

351

Note: The sample size is 700, spanning from January 3 to November 30, 2005. The Skewness column reports the statistics of skewness. The Kurtosis column presents the excess kurtosis statistics. Q(26) is the LjungBox statistic with the lag length equal to 26, to test for autocorrelation. JB denotes the JarqueBera statistics of test for normality. The ARCH column sets out the LM test statistics for ARCH. ** indicates signicance at the 1% level.

3. Econometric methods This section briey outlines the econometric methods used in this paper. Consider a return series ri,t that is generated by ri;t li;t hi;t ei;t ; hi;t xi di e 2 i;t1
1=2

i 1; 2; . . . ; n;

1 2

hi hi;t1 ;

(or fatter distribution tails than the normal distribution). When ti becomes innite, the normal distribution results. The Sk-t distribution as described by (3) is potentially useful for modeling nancial time series which are mostly fattailed and often skewed. Using the volatility series hi,t obtained by estimating Eqs. (1)(3) the conditional covariance matrix of asset returns ri,t (i = 1, 2, . . ., n) may be expressed as: H t Dt Rt Dt ; where 0 p h1;t B B B 0 B Dt B B . B . @ . 0 Rt diagQt 0 p h2;t .. . .. .. 0 . . 0 . . . 0 p hn;t 1 C C C C C C C A 4

where li,t, the conditional mean, contains AR(p) terms of ri,t plus a constant,4 such that the demeaned return series ri,t li,t will have iid standardized residuals ei,t after the 1=2 conditional volatility hi;t is ltered. For the two stock return series, the order p is chosen to be zero, while for the T-bond return series, p is set to 3, as the corresponding LjungBox Q(26) statistics become insignicant at the 5% level for the respective demeaned return series. Instead of introducing asymmetry to the GARCH model (Eq. (2)) for hi,t while assuming normality on ei,t, we work on the assumption that ei,t is driven by the Sk-t (skewed Student-t) distribution; that is, possible asymmetry occurs in ei,t rather than in hi,t. This practice follows closely the existing literature (see, e.g., Jondeau and Rockinger, 2005). The Sk-t distribution (pdf) is dened as: 8   ti 1 2 2 > > bi ci 1 bi ei;t ai =1ki > ti 2 > > > > > < if ei;t < ai =bi ; 3 Sk-tei;t jti ; ki   ti 1 > > > b c 1 bi ei;t ai =1ki 2 2 > i i > ti 2 > > > : if ei;t P ai =bi ; q t where ai  4ki ci tii 2 ; bi  1 3k2 a2 and ci C ti 1 = i i 2 p ti 1 pti 2C 2 . In Eq. (3), the parameter ki introduces asymmetry to the standard Student-t distribution: If ki = 0, then ai = 0 and bi = 1, giving rise to the standard Student-t distribution which is symmetric. If 1 < ki < 0 (0 < ki < 1), the distribution Sk-t(ei,tjti, ki) is negatively (positively) skewed. The parameter ti (>2) measures the degree of freedom. A nite ti means positive excess kurtosis
We have also explored the possibility that the 2004 austerity had an impact on the conditional mean by adding a dummy to li,t. The test results, however, show no signicant evidence of structural breaks in the level of average returns on T-bonds, Shanghai stocks and Shenzhen stocks. The results are available from the authors upon request.
4

and

Qt diagQt ;

where Rt is the conditional correlation matrix for standardized return residuals {ei}t (i = 1, 2, . . ., n) the elements in which, qij,t, represent conditional correlation between ei,t and ej,t. Qt may be taken as the conditional covariance matrix of the vector et, the elements in which, qij,t, can be used p to compute qij;t qij;t = qii;t qjj;t . To model Qt for estimating Rt, we adopt the diagonal version of the asymmetric DCC model proposed by Sheppard (2002). The model allows for dierent dynamics across correlations between dierent assets. With this model, we can explore the possibility that the correlations dynamic evolution of, say, the T-bond market vs. the Shanghai stock market is dierent from that for, say, the Shanghai stock market vs. the Shenzhen stock market. Moreover, this model also allows for asymmetries in the responses of correlations to information shocks (such as joint good and/or bad news). The diagonal version of the asymmetric DCC model is specied as: Qt Q A0 QA B0 QB G0 ZG A0 et1 e0t1 A B0 Qt1 B G0 zt1 z0t1 G; where 6

352

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

1 qe 1n B . C B e . C B q12 1 . C C for t 1; 2; . . . ; T ; QB C B . .. B . e qn1;n C . A @ . qe qe 1 1n n1;n 1 0 z qz 1 q12 1n B . C .. B z . C B q12 1 . C . C for t 1; 2; . . . ; T ; ZB C B . .. .. B . z . qn1;n C . A @ . qz qz 1 1n n1;n 1 1 0 0 b1 0 0 a1 0 0 B B . C . C .. .. B0 a B0 b . C . C . . C . . C 2 B B 2 AB C; B B C . .. .. C C B . B . .. .. @ . @ . . 0A . . 0A . . 0 0 an 0 0 bn 1 0 c1 0 0 B . C .. B0 c . C . . C B 2 and G B C: . .. .. C B . @ . . 0A . 0 1 qe 12 .. . .. . 0 0 cn In addition, et is a column vector of n standardized return residuals, while zt is a column vector containing all elements in et that are negative (or positive) and zeros otherwise. qe denotes the unconditional correlation coecient ij between ei,t and ej,t, and qz denotes the unconditional corij relation coecient between zi,t and zj,t. Notice that, when cs are set to zero, Eq. (6) collapses to a symmetric DCC model; and when ai = a(i = 1, 2, . . ., n), bi = b (i = 1, 2, . . ., n) and ci = c (i = 1, 2, . . ., n), Eq. (6) becomes a scalar version of the asymmetric DCC model which connes the dynamics to being identical across all the correlations between dierent assets. This restriction is unnecessary, and will limit the models applicability to dierent circumstances. To facilitate the exposition of our methodological innovations, we rewrite Eq. (6) in an element version as: h i ij qij;t 1 aij bij ij gij qz aij ei;t1 ej;t1 q bij qij;t1 gij zi;t1 zj;t1 ; 7

tional covariance between, say, the T-bond and Shanghai stock returns might be more responsive to bad news, whereas at the same time the conditional covariance between, say, the Shanghai and Shenzhen stock returns might be more responsive to good news. This issue has not yet received attention so far. To explore the possibility, we propose to modify the standard asymmetric DCC to allow for the coexistence of positive and negative leverage eects, and term such a modied model mixed asymmetric DCC (MADCC), the details of which are given as follows. Suppose that for i = 1 only and j = 1, . . ., n, asymmetry is negative-signed (i.e., zi;t z  Ieit < 0eit ; zj;t z  i;t j;t ij ij Iejt < 0ejt , and hence qz qz  EIeit < 0eit Iejt < 0ejt ), while for all other cases (i = 2, . . ., n and j = 1, 2, . . ., n) asymmetry is positive-signed (i.e., zi;t z  i;t ij ij Ieit > 0eit , zj;t z  Iejt > 0ejt , and hence qz  qz j;t EIeit > 0eit Iejt > 0ejt ). In this case, (7) becomes: h i ij qij;t 1 aij bij ij g qz aij ei;t1 ej;t1 q ij bij qij;t1 g z z ij i;t1 j;t1 for i = 1 and j = 1, . . ., n; and h i qij;t 1 aij bij ij g qz aij ei;t1 ej;t1 q ij  ij bij qij;t1 g z z ij i;t1 j;t1 for i = 2, . . ., n and j = 1, . . ., n. The resulting parameter restrictions are imposed in the following manner. When zi;t z  Ieit < 0eit and zj;t z  Iejt < 0ejt ; gij ci cj i;t j;t with i = 1 and j = 1, . . ., n are freely estimated to give the ^ij ci cj estimates as g ^^ , but gij = cicj with i = 2, . . ., n and j = 1, . . ., n are restricted to be zero. When zi;t z  i;t Ieit > 0eit and zj;t z  Iejt > 0ejt gij ci cj with i = 1 j;t and j = 1, . . ., n are restricted to be 0, but gij = cicj with i = 2, . . ., n and j = 1, . . ., n are freely estimated to give the ^ij ci cj estimates as g ^^ . We will subject these parameter restrictions, or our search for the particular MADCC, to statistical tests. We also consider possible structural breaks in the model. Take the case where breaks occur in the correlation mean as an example, and thus (6) becomes: Qt Q1 A0 Q1 A B0 Q1 B G0 Z 1 G Q2 A0 Q2 A B0 Q2 B G0 Z 2 G A0 et1 e0t1 A B0 Qt1 B G0 zt1 z0t1 G; 8

where aij = aiaj, bij = bibj, gij = cicj, zi,t, zj,t = I(eit < 0)eit, I(ejt < 0)ejt for negative-signed asymmetry (or zi,t, zj,t = I(eit > 0)eit, I(ejt > 0)ejt for positive-signed asymmetry), and i, j = 1, 2, . . ., n. Previous studies using asymmetric DCC models have ij restricted asymmetries (represented by gij qz and gij, zi,t1, zj,t1 in (7)) as extra responses to the same type of shocks, such as negative shocks, of all the correlations between assets i, j = 1, 2, . . ., n. However, it is possible that, in the Chinese context, the conditional covariance between dierent assets might have, concurrently, extra responses to news shocks of dierent types/signs. That is, the condi-

where Q1 is dened as Q for t < s (break date), and Q2 for t P s. Analogously Z 1 is dened as Z for t < s, and Z 2 for t P s. The null hypothesis is Q1 Q2 Q over the entire sample period, and it generates a value of the log likelihood function in (9). The alternative hypothesis is Q1 6 Q2 , and it generates another value of the log likelihood function. By comparing the two values of the log likelihood function using the likelihood ratio test statistic, we can decide on

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

353

whether or not the null can be rejected in favor of the alternative. Estimation of the models is carried out by maximizing the following log likelihood function of the DCC model with Sk-t distribution (see Jondeau and Rockinger, 2005): L
T X t1 T 1X lnSk-tet jt; k ln jH t gj 2 t1

Return volatility (T-bond)

0.9 0.8 0.7 0.6 0.5

0.4 0.3 0.2

with respect to the parameter vectors t, k and g (g contains all the parameters in the GARCH and DCC models). 4. Estimation and testing results Table 2 reports the estimation results for the Skew-tGARCH model for the three return series. All the d and h parameters for the conditional variances of the three return series are positive and statistically signicant at the 1% level, indicating volatility clustering of the three return series. Overall, volatility shocks appear to be persistent as d + h is close to unity for all the return series. Recall that the parameter ki introduces asymmetry to the standard Student-tdistribution. For the T-bond index, a negative k suggests that its return distribution is negatively skewed. On the other hand, a positive k indicates that the return distribution is positively skewed for the Shanghai and Shenzhen stock indices. The estimates of the parameter t are all nite and range between 4 and 11, albeit with a low precision for the two stock indices, signifying that the three asset returns are all driven by non-normal, fat-tailed distributions. Figs. 46 plot three series of return volatility from the Sk-t-GARCH model. There are clear dierences in the volatility pattern between the T-bond and the two stock return series. Overall, the T-bond volatility is relatively smaller than the stock volatilities, consistent with the wisdom that T-bonds have lower risk, or are safer, than stocks. The two stock volatilities are quite similar in terms of their patterns
Table 2 Estimation results of the Sk-t-GARCH model Parameter x d h t k T-bond market 0.0003** (1531) 0.2307** (50.67) 0.7693** (87.40) 4.6070** (11.17) 0.0588** (23.22) Shanghai stock market 0.0605** (10.77) 0.0440** (40.31) 0.9195** (254.73) 7.3712 (1.513) 0.1631** (54.13) Shenzhen stock market 0.0341** (11.63) 0.0429** (26.97) 0.9384** (251.1) 10.040 (0.651) 0.1064** (35.84)

0.1 0
/01/0 3 2003 /02 2003 /24 /04 2003 /03 /05 2003 /22 /07/0 1 2003 /08 2003 /08 /09 2003 /17 /11/0 3 2003 /12 2004 /11 /02/0 2 2004 /03 2004 /11 /04/2 0 2004 /06 2004 /04 /07 2004 /14 /08/2 3 2004 /09 2004 /30 /11 2004 /16 /12 2005 /24 /02/0 3 2005 /03 2005 /24 /05/1 0 2005 /06 2005 /17 /07/2 7 2005 /09 2005 /05 /10/2 0

2003

Fig. 4. Return volatility of Chinese T-bonds.

Return volatility (Shanghai stock)

5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0
/01/0 3 2003 /02/2 4 2003 /04/0 3 2003 /05/2 2 2003 /07/0 1 2003 /08/0 8 2003 /09/1 7 2003 /11/0 3 2003 /12/1 1 2004 /02/0 2 2004 /03/1 1 2004 /04/2 0 2004 /06/0 4 2004 /07/1 4 2004 /08/2 3 2004 /09/3 0 2004 /11/1 6 2004 /12/2 4 2005 /02/0 3 2005 /03/2 4 2005 /05/1 0 2005 /06/1 7 2005 /07/2 7 2005 /09/0 5 2005 /10/2 0

2003

Fig. 5. Return volatility of Shanghai stocks.

Return volatility (Shenzhen stock)

5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0
3/01 /03 200 3/02 /24 200 3/04 /03 200 3/05 /22 200 3/07 /01 200 3/08 /08 200 3/09 /17 200 3/11 /03 200 3/12 /11 200 4/02 /02 200 4/03 /11 200 4/04 /20 200 4/06 /04 200 4/07 /14 200 4/08 /23 200 4/09 /30 200 4/11 /16 200 4/12 /24 200 5/02 /03 200 5/03 /24 200 5/05 /10 200 5/06 /17 200 5/07 /27 200 5/09 /05 200 5/10 /20

200

Fig. 6. Return volatility of Shenzhen stocks.

Note: This table gives the estimates of the parameters in Eqs. (2) and (3). When carrying out estimation, li,t in Eq. (1) is a constant for the two stock returns, but equals a constant plus the AR(1), AR(2) and AR(3) terms of ri,t for the T-bond returns, such that the demeaned three return series 1=2 ri,t li,t have iid standardized errors ei,t after the conditional volatility hi;t is ltered. In parentheses are the t-statistics. ** indicates signicance at the 1% level.

and magnitudes. There is a notable observation that the Tbond volatility exhibits spikes far above the average amplitude around May of 2004 (the peak value, 0.8132, occurred on May 10, 2004). The spikes could be an indication that the policy shocks were large enough to have caused breaks in the correlation of the T-bond and stock markets. Whether this is the case, however, needs to be formally tested, to which we now turn.

354

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359 Note: NADCC stands for the negative asymmetric DCC model. PADCC stands for the positive asymmetric DCC model. MADCC stands for the mixed asymmetric DCC model. LRT stands for the log likelihood ratio test statistics. In parentheses are the t-statistics. In brackets are the p values. **indicates signicance at the 1% level. * indicates signicance at the 5% level. 0.0504** (3.0393) 0.4069** (42.424) 0.3217** (25.474) 0.0235 0.7870 (0.3132) (0.3104) Shanghai stock 0.2086** 0.7914** (19.712) (81.680) Shenzhen stock 0.1936** 0.6886** (29.199) (36.965) H0: NADCC is true; H1: MADCC is true H0: PADCC is true; H1: MADCC is true H0: no break in mean; H1: two breaks in mean T-bond 0.0214** (8.5137) 0.0000 (0.0000) 0.1178 (1.5382) 0.0000 (0.0001) 0.0000 (0.0000) 0.0790 (0.6594) 0.8710** (5.7618) 0.5617** (5.2382) 0.5623** (9.7834) 0.0000 0.0219 (0.0000) (1.0705) 0.3659** 0.2235** (4.4068) (16.154) 0.3587** 0.1629** (16.257) (11.265) LRT = 28.0172** [0.0000] LRT = 18.7920** [0.0003] LRT = 26.788** [0.0002] 0.9276** (95.561) 0.2058* (2.1674) 0.2146 (1.6357)

We rst consider our conjecture that positive and negative leverage eects coexist in the correlation structure of the Chinese T-bond and stock markets. To do this, we compare, via some statistical tests, three versions of the asymmetric DCC: (a) the positive asymmetric DCC (PADCC) where the conditional covariance of standardized return residuals responds more strongly to all the residuals joint positive than negative shocks; (b) the negative asymmetric DCC (NADCC) where the conditional covariance of standardized return residuals responds more strongly to all the residuals joint negative than positive shocks; and (c) the mixed asymmetric DCC (MADCC) where the conditional covariance of standardized return residuals responds more strongly to some residuals joint positive shocks than these residuals joint negative shocks, and more strongly to other residuals joint negative shocks than these residuals joint positive shocks. The t-statistics associated with the estimates of the ci parameters in Table 3 convey interesting messages regarding the three versions of the asymmetric DCC model. For the NADCC, the c estimate associated with T-bond is statistically signicant at the 1% level, while the c estimates associated with Shanghai stock and Shenzhen stock are statistically insignicant even at the 10% level. The NADCC thus suggests that c1 = cT-bond > 0, c2 = cSH-stock = 0 and c3 = cSZ-stock = 0; that is, the conditional covariance of standardized return residuals has a greater response to negative than positive shocks to the T-bond market only. For the PADCC, the c estimate associated with Tbond is statistically insignicant even at the 10% level, while the c estimates associated with Shanghai stock and Shenzhen stock are statistically signicant at the 1% level (see Table 3). The PADCC thus suggests that c1 = cT-bond = 0, c2 = cSH-stock > 0 and c3 = cSZ-stock > 0; that is, the conditional covariance of standardized return residuals has a greater response to all joint positive than negative shocks to the Shanghai and Shenzhen stock markets, but not to joint positive than negative shocks to the Tbond and Shanghai (or Shenzhen) stock markets. The fact that the NADCC and PADCC can only capture either one, but not all, of the above-discussed two facets of asymmetry motivates our search for a particular MADCC that is capable of embracing both of these two facets in one model. Indeed, Table 3 shows that, for the MADCC, all the three c estimates now become statistically signicant at the 1% level. Moreover, the likelihood ratio test statistics signicant at a higher than 1% level enable us to reject the NADCC and the PADCC in favor of the MADCC. The MADCC suggests that c1 = cT-bond > 0 only when shocks to the T-bond market is negative; c2 = cSH-stock > 0 and c3 = cSZ-stock > 0, only when shocks to the Shanghai and/or Shenzhen stock markets are positive. That is, the conditional covariance of standardized return residuals responds more strongly to negative than positive shocks to the T-bond market, but more strongly to joint positive than negative shocks to the Shanghai

MADCC PADCC Table 3 Estimation results of dierent asymmetric DCC models NADCC

ai

bi

ci

ai

bi

ci

ai

bi

ci

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

355

and/or Shenzhen stock markets. These results provide evidence in support of the coexistence of positive and negative leverage eects in the Chinese capital markets. Next, consider possible changes in the correlation structure. Since the MADCC model outperforms the NADCC and PADCC models, we report the test results only for the MADCC model. The tests were performed by searching break points over the period between April 2004 and December 2004. Table 3 sets out the likelihood ratio statistic for testing the null hypothesis of no break in the correlation mean against the alternative hypothesis of two breaks in the correlation mean. The result shows that the null is rejected at a higher than 1% level in favor of the alternative. In fact, we also tested the null against the alternative of one break in the correlation mean over the period from April to December, 2004, and the values of the log likelihood function associated with one break in the correlation mean were either smaller than or very close to that with no break. In addition, it is possible that the two structural breaks also exist in the dynamics of correlation (i.e., in the A and B matrixes). In order to see if this is the case, we again used the log likelihood functions to compare between breaks in the correlation mean only and breaks in both the correlation mean and dynamics. There were no signicant improvements in the log likelihood functions that could allow us to reject the null of two breaks in the correlation mean only in favor of the alternative of two breaks in both the correlation mean and dynamics; and this applies to the NADCC, PADCC and MADCC models. To sum up, we conclude that the MADCC model with two breaks in the correlation mean is the most appropriate model to capture the possibility that there coexist negative and positive leverage eects, and two structural breaks, in the correlation of the Chinese T-bond and stock markets. Therefore, the estimation results of the model are used in aid of plotting a number of gures that can provide intuitive insights into the main issues investigated in this paper. 5. The impacts of policy and news shocks on correlation Figs. 79 depict the dynamics of correlation, derived from the MADCC model, between, respectively, the Tbond and Shanghai stock returns, the T-bond and Shenzhen stock returns, and the Shanghai and Shenzhen stock returns. Recall from the preceding section that two structural breaks were detected by the MADCC model. The two breaks can be clearly seen by an inspection of Figs. 79. Several messages emerge from the three gures, and are relevant to the questions posed earlier in Section 1. First, a series of more austere macroeconomic contractions brought into eect in April 2004 started to be felt concretely by the domestic nancial markets in May 2004. This provides evidence that the large policy shocks did have certain impacts on nancial markets albeit with a lag. The spikes in the volatility of the T-bond returns as shown in Fig. 4 are indeed an indication of structural change in the

Correlation (T-bond vs Shanghai stock)

0.2 0.15 0.1 0.05 0 -0.05 -0.1
2003 /01/0 3 2003 /02/2 4 2003 /04/0 3 2003 /05/2 2 2003 /07/0 1 2003 /08/0 8 2003 /09/1 7 2003 /11/0 3 2003 /12/1 1 2004 /02/0 2 2004 /03/1 1 2004 /04/2 0 2004 /06/0 4 2004 /07/1 4 2004 /08/2 3 2004 /09/3 0 2004 /11/1 6 2004 /12/2 4 2005 /02/0 3 2005 /03/2 4 2005 /05/1 0 2005 /06/1 7 2005 /07/2 7 2005 /09/0 5 2005 /10/2 0

Fig. 7. Plot of the conditional correlation between the T-bond and Shanghai stock returns.

Correlation (T-bond vs Shenzhen stock)

0.25 0.2 0.15 0.1 0.05 0 -0.05 -0.1
2003 /01/0 3 2003 /02/2 4 2003 /04/0 3 2003 /05/2 2 2003 /07/0 1 2003 /08/0 8 2003 /09/1 7 2003 /11/0 3 2003 /12/1 1 2004 /02/0 2 2004 /03/1 1 2004 /04/2 0 2004 /06/0 4 2004 /07/1 4 2004 /08/2 3 2004 /09/3 0 2004 /11/1 6 2004 /12/2 4 2005 /02/0 3 2005 /03/2 4 2005 /05/1 0 2005 /06/1 7 2005 /07/2 7 2005 /09/0 5 2005 /10/2 0

Fig. 8. Plot of the conditional correlation between the T-bond and Shenzhen stock returns.

Correlation (Shanghai stock vs Shenzhen stock)

1 0.98 0.96 0.94 0.92 0.9 0.88 0.86 0.84 0.82
2003 /01/0 3 2003 /02/2 4 2003 /04/0 3 2003 /05/2 2 2003 /07/0 1 2003 /08/0 8 2003 /09/1 7 2003 /11/0 3 2003 /12/1 1 2004 /02/0 2 2004 /03/1 1 2004 /04/2 0 2004 /06/0 4 2004 /07/1 4 2004 /08/2 3 2004 /09/3 0 2004 /11/1 6 2004 /12/2 4 2005 /02/0 3 2005 /03/2 4 2005 /05/1 0 2005 /06/1 7 2005 /07/2 7 2005 /09/0 5 2005 /10/2 0

Fig. 9. Plot of the conditional correlation between the Shanghai and Shenzhen stock returns.

dynamics of correlation: they were accompanied by an abrupt increase in the correlations between the T-bond and two stock returns as shown in Figs. 7 and 8, and in the correlation between the two stock returns as shown in Fig. 9. In particular, the rises in the two bondstock correlations are shown to be signicantly greater than that in the stockstock correlation, approximately at a ratio of 0.15 0.02. This is consistent with the observation that the Tbond volatility underwent relatively large spikes whereas the stock volatilities did not, around May 2004, implying that the T-bond market is more susceptible to the macroeconomic contractions than the stock markets. The rise

356

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

in the correlation increased the overall risk of the portfolios that include T-bonds and stocks, as the probability that these assets prices rise/fall together was increased. Second, the impact of the April austerity programs on the correlation lasted for about 5 months, as Figs. 79 show that the rst break date is May 14, 2004 while the second is October 29, 2004, although the second break point in the stockstock correlation is indiscernible. The news media reported, using data released a couple of months later, that the macroeconomic austerity measures turned out to be eective and successful in three main aspects: the growth of xed investment had declined rapidly; the money supply had fallen to the targeted range set by the central bank; and the upward trend of price movements had been stopped or even slightly reversed. Among economists and the economic authorities, consensus was then reached that macroeconomic stabilization should now switch from ghting against overheating/ination to preventing potential overcooling/deation. This does not mean a switch from drastic contractions to comprehensive expansions. In late 2004, in fact, there was a call for moderate macroeconomic policies that would lessen the tightness of the existing austerity policies. The above may explain 5-month duration of the policy impact on the correlation of the domestic nancial markets. Third, before the rst structural break, the two bond stock correlation series were positive but took negligible values around 0.025. After the second break, the correlation series became negative with a slightly greater average value of 0.05 (in absolute terms). The two observations, low stockbond correlations and change of the sign of the correlations, are interpreted as follows. We propose that the observed low bondstock correlations should be due mainly to the low level of stockbond market integration, for two reasons. First,5 recall from Eq. (6) that correlations investigated are between standardized errors eit, that is, between returns which are demeaned as well as volatility-ltered (i.e, mean- as well as riskadjusted). This suggests that the dierence in the riskreturn characteristics between stocks and bonds is unable to account for the observed low bondstock correlations. Therefore, that information shocks (represented by eit which are iid) to the stock and bond markets are very weakly correlated could be attributed to the two markets being still highly segmented. Serious underdevelopment of the bond market relative to that of the stock market in China, as noted by XIANG Jun the Deputy Governor of the Peoples Bank of China (http://www.china138.com/news/show.asp?id=46052), is an important reason for this segmentation. Second, as reviewed in Section 1, previous studies (Kim et al., 2006; Cappiello et al., 2003) have reported quite large estimates of the interbondstock correlations for European markets especially

prior to the advent of EMU. Note that Cappiello et al. (2003) also use mean- and risk-adjusted returns for estimating their conditional correlations. Compared with those European results, our Chinese results can be taken to imply that the integration level of Chinas T-bond and stock markets is still low, at least relative to that for the European counterparts at some points in time. The negative bondstock return correlations following the second break could be a signal that the ight-to-quality phenomenon has started to characterize the behavior of the Chinese investors, albeit still to a small degree. In addition to the dynamics of correlation, we also examine asymmetry in the responses of correlation to joint bad news and joint good news. Following Kroner and Ng (1998), we employ the news impact surfaces. If the correlation between two asset returns has a greater/smaller response to joint bad news than to joint good news, the news impact surface will delineate intuitively such correlation asymmetry. And it is more relevant to examine asymmetry related to the same-signed information shocks than that related to the opposite-signed information shocks. The reason, as stated in Section 1, is because asymmetry for the same-signed shocks has more important nancial implications than for the opposite-signed shocks. A news impact surface for correlation is constructed using all the estimated parameters of a DCC model.6 Thus, dierent DCC models employed would lead to dierent news impact surfaces. Our news impact surfaces are based on the parameter estimates of the MADCC model7 with opposite-signed (i.e., both positive-signed and negativesigned) correlation asymmetries, while using the original KronerNg news impact surfaces would involve the parameter estimates of a DCC model with the same-signed (i.e., either positive-signed or negative-signed) correlation asymmetry. An application of the KronerNg news impact surfaces for correlation which have greater response to joint bad news than to joint good news (i.e., negativesigned correlation asymmetry) can be found in Cappiello et al. (2003). These three authors provided the exact formulae for calculating the news impact surfaces based on what we now call the NADCC model. Compared with their formulae, our MADCC-based news impact surface formulae seem more complex, a price paid for complexity. However, the advantage of our new technology is worth the price. In the Chinese context, for example, using the original KronerNg news impact surfaces would be unable to delineate positive-signed correlation asymmetry if based on the NADCC, or negative-signed correlation asymmetry if based on the PADCC. Our MADCC-based news impact

5 We thank a referee for the suggestion to conduct this risk-adjusted analysis.

Kroner and Ng (1998) introduced the notion of the news impact surfaces for, respectively, covariance, conditional variance and correlation. In this paper, we are only interested in the news impact surfaces for correlation. 7 The formulae for calculating the news impact surfaces using the parameter estimates of the MADCC model are available from the authors upon request.

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

357

surfaces enable one to view all co-existent correlation asymmetries, negative-signed and positive-signed, in the proper perspective. Figs. 1012 depict three news impact surfaces (each having two panels (a) and (b) taken from two dierent angles) for three correlations, respectively: The T-bond vs. Shanghai stock correlation, the T-bond vs. Shenzhen stock correlation, and the Shanghai vs. Shenzhen stock correlation. One can see that the three correlation news impact surfaces are highly asymmetric. Figs. 10 and 11 shows a larger response to joint negative shocks (in the , quadrant) than to joint positive shocks (in the +, + quadrant), of the two bondstock correlations. In other words, the likelihood that T-bonds and Shanghai (or Shenzhen) stocks tend to fall together given that they have already fallen together is greater than the likelihood that they tend to rise together given that they have already risen together. The economic intuition of the result is this. After both the T-bond and stock markets have been hit by bad news (i.e., joint negative shocks), a majority of investors may respond by selling both T-bonds and stocks, leading both the T-bond and stock prices to further fall together. On the other hand, after both the T-bond and stock markets have been hit by good news (i.e., joint positive shocks), a majority of investors may shift funds from risky stocks to relatively safe T-bonds, causing a further and large rise in the T-bond price but a small or no rise or even a fall in the stock price. So, bondstock return correlations in the case of joint good news are smaller than in the case of joint bad news. Fig. 12, however, marks asymmetry for the stockstock correlation opposite to those for the bondstock correlations. It shows a larger response to joint positive shocks (in the +, + quadrant) than to joint negative shocks (in the , quadrant), of the stockstock correlation. Put dierently, the probability that Shanghai and Shenzhen stocks tend to rise together given that they have already

risen together is greater than the probability that they tend to fall together given that they have already fallen together. We propose to interpret this nding as follows. After both the two stock markets have been hit by good news (i.e., joint positive shocks), more investors respond by purchasing both Shanghai and Shenzhen stocks than not doing so, leading both the two stock prices to further rise together. On the other hand, after both the two stock markets have been hit by bad news (i.e., joint negative shocks), more investors do not hurry to sell than those do. These are the economic intuition for the nding that the probability of rising together following rising together is greater than the probability of falling together following falling together. It is worthwhile to devote more space to the stories behind the observation that the stockstock correlation is greater following joint positive shocks than joint negative shocks. According to Li (2004, p. 326), under information asymmetry, the Chinese fund managers are risk-lovers. When a stock market turns bullish, they tend to be panicked into purchasing stocks, in the expectation that the stock prices will rise further. When the stock market turns bearish, however, they are less panicked into selling stocks, in the hope that market downturns will not last for long. These fund managers are aware of the possible existence and eventual burst of bubbles, but also subjectively believe that they will not encounter the bust of bubbles while their subsequent followers will. In addition to this psychological bias, there also exist distorted incentives: When an investment is making money, the fund managers share the prots (and the resultant performance-based remuneration is quite high), but when an investment is losing money, only the fund holders are the bearers of the losses. The psychological bias and the distorted incentives thus lead to the asymmetric behavior of the Chinese fund managers across market upturns and market downturns. Li (2004) shows that such asymmetric behavior is not conned to fund

Fig. 10. The correlation news impact surface for the T-bond and Shanghai stock returns.

358

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

Fig. 11. The correlation news impact surface for the T-bond and Shenzhen stock returns.

Fig. 12. The correlation news impact surface for the Shanghai and Shenzhen stock returns.

managers: Individual investors are also found to be characterized by this behavior. Moreover, researchers have found a signicant herding eect in China, even more signicant than in the US [see Li (2004, p. 325), for the review of the studies on the herding eect in Chinas stock markets]. It may be this herding eect that helps the asymmetric movements of individual markets lead to the asymmetric correlation between stock markets in China. 6. Conclusions We now summarize the main ndings of this study that are believed to have important implications for policymakers, market participants and investors in China. The co-movements of the Chinese capital markets do react to large macroeconomic policy shocks as evidenced by structural breaks in the correlation following the drastic 2004 macroeconomic austerity measures whose impact

lasted for about 5 months. The T-bond market and its correlations with the Shanghai and Shenzhen stock markets are found to have borne more of the brunt of the macroeconomic contractions than the two stock markets and their correlation. Overall, however, the level of Chinas bondstock market integration is still low, at least lower than that in Europe at some points in time, although Chinas stockstock market integration has reached a quite high level. In addition, the relatively small volatility in Tbond returns implies that investors could reap diversication benets via ight to quality (i.e., by moving their capital out of riskier equities and into safer government securities). Such a portfolio-rebalancing strategy seems to have started to be considered by Chinese investors after the episode of macroeconomic contractions in 2004. As a methodological innovation, we propose a new version of the asymmetric DCC model, MADCC, which seems to have been quite helpful, as it enables us

X.-M. Li, L.-P. Zou / Journal of Banking & Finance 32 (2008) 347359

359

to successfully capture the coexistence of positive-signed and negative-signed asymmetries in the correlation structure, a unique characteristic of the Chinese capital markets. The news impact surfaces based on the MADCC provide intuitionalized evidence that the bondstock correlations tend to increase only when their returns have both been hit by bad news, but the stockstock correlations tend to increase only when their returns have both been hit by good news. Acknowledgements We thank Henk Berkman, Charles Corrado, Ben Jacobsen, participants at the 14th SFM conference, the 2006 AFS Meeting and the Department of Commerce Brown Bag Seminar at Massey University, and two anonymous referees for their helpful comments and suggestions. The usual disclaimers apply. References
Barsky, R.B., 1989. Why dont the prices of stocks and bonds move together? American Economic Review 79, 11321145. Berben, R.P., Jansen, W.J., 2005. Bond market and stock market integration in Europe. DNB Working Papers 060, Netherlands Central Bank, Research Department. Cappiello, L., Engle, R., Sheppard, K., 2003. Asymmetric dynamics in the correlations of global equity and bond returns. Working paper No. 204. European Central Bank. Chordia, T., Sarkar, A., Subrahmanyam, A., 2005. An empirical analysis of stock and bond market liquidity. Review of Financial Studies 18, 85129.

Connolly, R., Stivers, C., Sun, L., 2005. Stock market uncertainty and the stockbond return relation. Journal of Financial and Quantitative Analysis 40, 161194. Connolly, R., Stivers, C., Sun, L., 2007. Commonality in the timevariation of stockbond and stockstock return comovements. Journal of Financial Markets 10, 192218. Ilmanen, A., 2003. Stockbond correlations. Journal of Fixed Income 12, 5566. Jondeau, E., Rockinger, M., 2005. Conditional asset allocation under nonnormality: How costly is the mean-variance criterion? Research paper No. 132. FAME International Centre for Financial Asset Management and Engineering, February, 2005. Kim, S.J., Moshirian, F., Wu, E., 2005. Dynamic stock market integration driven by the European Monetary Union: An empirical analysis. Journal of Banking and Finance 29, 24752502. Kim, S.J., Moshirian, F., Wu, E., 2006. Evolution of international stock and bond market integration: Inuence of the European Monetary Union. Journal of Banking and Finance 30, 15071534. Kroner, K., Ng, V., 1998. Modelling asymmetric comovements of asset returns. Review of Financial Studies 11, 817844. Li, L., 2002. Macroeconomic factors and the correlation of stock and bond returns. Yale ICF working paper No. 02-46. Li, X.-M., 2003a. Time-varying informational eciency in Chinas Ashare and B-share markets. Journal of Chinese Economic and Business Studies 1, 3356. Li, X.-M., 2003b. China: Further evidence on the evolution of stock markets in transition economies. Scottish Journal of Political Economy 50, 341358. Li, X.-D., 2004. Behavioral Finance: Theory and Evidence from Chinese Stock Market. Shanghai Sanlian Shudian Publishing House, Shanghai. Scruggs, J.T., Glabadanidis, P., 2003. Risk premia and the dynamic covariance between stock and bond returns. Journal of Financial and Quantitative Analysis 38, 295316. Sheppard, K., 2002. Understanding the dynamics of equity covariance. Manuscript, UCSD. Yin, J., 2005. The linkage between nancial markets in China: 20002004. Forum for Economic Development, Working Paper No. FC20050037.